A machine learning portfolio across telecom analytics (Sofrecom, Orange Group, Tunisia Telecom), forecasting, NLP, fraud detection, segmentation, LLM pipelines, graph analytics, VC analytics (Kiota Intelligence), and production AI — built across Zindi competitions, client engagements, and freelance projects.
Filter by company to explore the full portfolio.
Intelligent filtering and classification for telecom infrastructure imagery. Multi-stage pipeline: utility detection → photo/plan segmentation → infrastructure type identification. Custom LLM prompts, base64 encoding, BigQuery integration, confidence scoring and automated routing.
Weekly multi-horizon forecasting for 4 KPI targets across thousands of IDs. Semi-automatic pipeline: Step 0 meta-search → lag search → feature selection → hyperparameter tuning. Multi-seed robustness, dead-ID routing, BigQuery integration, rolling backtest.
Automated extraction system for audit reports using GPT-4. Multi-page PDF processing with image conversion, Pydantic data validation, modular Python architecture, automated JSON output generation.
12-month forecast (weekly/monthly) for global, departmental, and offer-specific dimensions. Limited historical data. Achieved ~5% RMSE through ensemble modeling using R's feasts and fable. Migrated to Dataiku production.
Converted Python scripts to production-grade Dataiku workflows. Prioritized SQL and native Dataiku components. Validated MLForecast function effectiveness, tested additional predictive variables, confirmed solution robustness.
K-means behavioral segmentation from Analytics Base Table built from multiple data sources. Two phases: unsupervised clustering to discover segments, then supervised classification to assign new customers. Delivered actionable insights for CVM strategy.
The hardest part was building the label — no reliable ground truth existed. Built the target from scratch: behavioral hypotheses → unsupervised discovery → cluster profiling → rule extraction → supervised monthly scoring system.
Detected household-level ties inside broader telecom communication communities. Used ranked contact behavior, shared cell-tower usage, and graph partitioning to distinguish households from extended social circles.
CVM performance dashboards, network quality KPIs, data service penetration, customer segmentation profiles, sales analytics and market growth monitoring. VBA automation for PowerPoint reporting.
Trajectory-aware pipeline forecasting cumulative COVID-19 deaths for ~180 countries. Classifies each country by trajectory (UP, DOWN, NEAR_ZERO, REST), routes to matched forecasting strategy. Ranked 4th — only 47 succeeded among 884.
Detecting fraudulent meter manipulation for Tunisian utility STEG (200M Dinars lost). 15 years of billing history, 5.6% positive rate. Stochastic feature selection across 100 LightGBM iterations, 14-model diversity ensemble. Ranked 6th / 191.
Predict when and where road incidents will occur next in Cape Town. Top 11%, 13th position — only 113 succeeded to submit among 711 competitors.
Predict which individuals are most likely to have or use a bank account. Top 13% — 678 submitted out of 1288.
Weekly wholesale maize price prediction for 5 Kenya counties, 2 weeks ahead. Ensemble of LightGBM, XGBoost, ARIMA, ETS and Prophet.
Predicted support ticket escalation before it occurs using structural features: response latency, re-open rate, message length progression. Structural signals outperformed raw text sentiment.
Predicted ambulance response times using geospatial features, dispatch queue signals, and temporal patterns. True bottleneck was pre-dispatch wait, not road distance.
Predict the estimated time of arrival for motorbike deliveries in Nairobi. Top 26% — 431 submitted out of 1143.
Forecast weekly sea turtle rescues for Local Ocean Conservation in Kenya. Seasonal pattern extraction, weather and oceanographic feature engineering.
Classify text from university students towards a mental health chatbot. Top 36% — 492 submitted out of 900.
Multi-modal prediction combining tabular features with Arabic and English NLP signals. Custom feature engineering for order patterns, menu diversity, behavioral recency.
Model energy consumption for 5G network infrastructure. Feature engineering on network configuration and usage patterns.
Predict rate of households below a certain wage at each ward in South Africa. Top 38% — 258 submitted out of 498.
Predict soil humidity in 5-minute increments using sensor data from low-cost DIY IoT devices. Top 37%.
Predict flood extent caused by storms in southern Malawi. Geospatial features, environmental data.
Predict early childhood development outcomes from survey and socio-demographic data.
Predict engagement patterns of new users on the Zindi platform.
Predictive models for the VC industry: startup lifetime prediction, Series A funding probability, pre-valuation estimation. Built entirely in R.
Created a matching system to select the right investors for a particular entrepreneur based on different filters. Hypothesis testing for funding round variations.
Built Shiny dashboard for funding visualization, dynamic filtering tools, multidimensional outlier detection, temporal clustering analysis. Automated RMarkdown reports with CSS styling.
Developed scalable text classification model with automated category prediction pipeline. Framework for handling future data classification.
From very sensitive sensors (millisecond-level data), predict the correct weight or reduce maximum error between real value and predicted one. High-precision regression from IoT data.
Designed comprehensive ML curriculum and hands-on labs: Linear/Logistic Regression, SVM, Decision Trees, K-means. Python notebooks covering supervised and unsupervised learning.
Facebook API integration for metrics collection. Sentiment analysis on user comments. Actionable insights for social media strategy.
Developed behavioral pattern recognition for homeowner identification, real-time intrusion detection, automated alarm management system. Smartfox Project with MyFox. Published research book (EUE).